648 Machine Learning Predicts Cavernous Sinus Invasion of Pituitary Adenomas

Abstract

INTRODUCTION: Cavernous sinus invasion (CSI) of pituitary adenoma can lead to incomplete resection, failure of endocrinological remission, and a higher tumor recurrence rate. Therefore, precise approaches toward CSI may help guide treatment strategies and patient counseling. METHODS: 135 patients (median age = 50.14±15.56; 1.05:1 male to female ratio) with pathologically confirmed pituitary adenoma were retrospectively identified. High-resolution, coronal T1-contrast enhanced pituitary MR images were used. The study cohort was separated into training (n=101) and testing (n = 34) sets. Bounding box applied over the pituitary gland-cavernous sinus complex served to generate input image data. A 2D ResNet50V2 convolutional neural network was used as the base architecture for model development. Intraoperative inspection of the cavernous sinus and histopathologic analysis of the cavernous sinus medial wall served as ground truths for CSI. RESULTS: 'The receiving operating characteristic (ROC) and precision-recall (PR) curves were used to assess model performance. The model's average area under (AU) ROC was 78.6% compared to Knosp AU ROC %64.7. The model AU PR curve was 82.7% compared to Knosp AU PR curve of 76.7%. The sensitivity and F1 score of the model were 71% and 0.73, respectively, compared to Knosp sensitivity and F1 scores of %35 and 0.50. CONCLUSIONS: Inherent challenges exist regarding the pre-operative evaluation of CSI of pituitary adenoma on MRI by human visual inspection alone. Our study demonstrates the potential for MRI-based machine learning analysis of CSI of pituitary adenomas. A future multi-site study that explores model generalizability could further examine the utility of AI integration in the clinical workflow.